Cannes: OpenAI Gets into Ads and Meta into Ad Automation

Cannes (I): OpenAI Debuts in Cannes and Becomes an Ad Agency

OpenAI made its debut at the Cannes Lions, not from a Croisette rooftop terrace, but on a semi-secluded patio a 15-minute walk from the festival center. Dave Dugan, Head of Global Ads Solutions and a twelve-year Meta veteran, announced “a completely new advertising product.” The numbers behind it: 900 million weekly users, with about 20 percent of queries having direct commercial intent. According to Axios, OpenAI has told investors it expects to reach $100 billion in revenue in four years, a milestone that took Meta 17 years. CRO Denise Dresser doesn't want to measure success by impressions, but by whether the advertising “helps people get things done.” Advertising is intended to subsidize free access, and conversational data will never be shared with advertisers. With the coding agent Codex, creative directors without programming skills will be able to build their own production workflows. → AI Secret

Synthszr Take: There's an ex-Meta manager standing on a terrace, explaining that ChatGPT is the opposite of scrolling while importing Meta's business model. For years, Sam Altman avoided advertising for what it is: the moment when the user is no longer the customer, but the product. Now it's here, neatly packaged as the “Intelligence Economy” and as a mission to give more people access. The $100 billion in four years is a statement against its own computing costs, because the 900 million free users consume a fortune that someone has to pay for. The guardrail question will be interesting: “We never share conversational data” sounds good, but the entire value of ChatGPT advertising is precisely this intent that the user reveals in conversation. Back in March, it was said here that pure scaling does not lead to AGI; the advertising business is the honest answer to how to finance scaling until then. Anyone using ChatGPT should now assume that every travel plan to the Alps is an upper-funnel signal.

Cannes (I): Meta Automates Advertising and WPP Loves It

In Cannes, Meta unveiled a new generation of AI tools for advertisers and agencies designed to close the so-called 'closed loop': a system that identifies successful ad creatives, analyzes their impact, and independently generates new variations from them. Nicola Mendelsohn, Head of Global Business Group, describes it as a tool that learns a brand's identity and tonality from existing ads and uses these insights for generative ad creative creation. WPP will be the first agency partner to test the solution, integrated directly into WPP Open, the group-wide AI platform that combines planning, creation, production, and media management in one system. WPP had only just announced a deeper collaboration with Adobe in the spring and is centrally integrating partners like AWS and Meta. At the festival, there was much talk of the transition from the “search-and-click web” to a conversational internet experience, in which platforms like Meta, Pinterest, or OpenAI are positioning themselves as recommendation engines. Critics warn that such systems primarily reproduce what has worked in the past, making campaigns more similar. In contrast, the creative agency Fifty Thousand Feet emphasizes imagination and cultural understanding as the crucial differentiating factors. → MEEDIA Daily Update

Synthszr Take: The closed loop optimizes each individual campaign, and the advertising landscape as a whole gravitates toward the same center. This is exactly what the Milan restaurant experiment showed: as soon as ChatGPT was gone and the restaurateurs wrote themselves, lexical diversity increased by 15 percent. The same tool that makes life easier for each individual makes everyone interchangeable. Efficiency on the micro level, entropy on the macro level. Meta's system learns what performed in the past and delivers exactly that back: best practice is the new mediocrity, professional and passionless. For WPP, this means outsourcing the production machine to Meta and asking the real question of what's left when everyone works with the same stack and the same AI output. Taste becomes the last moat, and it can't be generated in a pipeline. In this world, anyone who just produces the same thing faster has already lost the competition before the first ad creative is served.

OpenAI is Building Its Own Chip, Becoming More Independent from Nvidia

OpenAI and Broadcom have unveiled Jalapeño, an ASIC built specifically for inference that, according to both companies, is set to deliver better performance per watt than current top-tier hardware. The chip is not a repurposed training accelerator but is designed from the ground up for the behavior of large language models and upcoming agentic workloads: high throughput with low latency, a massive compute-chiplet with six HBM modules instead of cheaper DRAM. The estimate from the wafer image suggests a die size of around 840 mm², dangerously close to the 858 mm² reticle limit of EUV lithography. Engineering samples are already running in the lab and handling ML workloads like GPT-5.3-Codex-Spark. The pace is remarkable: tape-out in nine months, with deployment starting in late 2026. Specific benchmarks, memory configuration, and performance targets remain undisclosed, so the efficiency promises should be taken with a grain of salt. Whether Jalapeño can compete with Nvidia's Rubin and AMD's MI400 is the real open question, not the comparison with the already aging Blackwell generation. → www.tomshardware.com

Synthszr Take: In February, OpenAI confessed that inference is getting brutally expensive. Jalapeño is the logical response: if you're paying per token and pushing billions of tokens a day, you eventually reclaim the margin through your own silicon. Nine months from idea to tape-out is a pace that makes you read twice if you're a hardware person, and it suggests that AI played a significant role in the chip design (otherwise, the cycle hardly pays off). The compute discipline is currently shifting from training to inference, and that's where the money is, because reasoning and agents make every query more expensive. The missing benchmarks are the real risk: an 840 mm² die near the reticle limit sounds muscular but says nothing about real-world utilization. Anyone betting on OpenAI as a platform should price in the vendor lock-in now, because custom hardware is the strongest lock-in point a foundation lab can build. The real excitement will come in late 2026, when the chips compete against Rubin, not just a press release.

Anthropic Accuses Alibaba of the Largest Distillation Campaign Against Claude to Date

Anthropic has informed US senators and the White House that operators associated with Alibaba's Qwen lab used nearly 25,000 fake accounts between April and June to scrape Claude's capabilities. The activity involved almost 29 million exchanges, specifically targeting software engineering and agentic reasoning—precisely the model's most commercially valuable skills. This makes Alibaba's volume alone greater than the three previous cases (DeepSeek, MiniMax, and Moonshot) combined, which involved 16 million exchanges and 24,000 fake accounts. Distillation means feeding a frontier model with prepared queries, collecting the responses, and using them to train a cheaper competing system. Alibaba's US depositary receipts fell over three percent to below $100; the Pentagon had placed the conglomerate on its blacklist of Chinese military companies on June 8, a move Alibaba is challenging in court this week. The situation is delicate: Anthropic is asking the government for help while simultaneously clashing with the Trump administration over export controls for its Fable-5 and Mythos-5 models. The company, valued at $965 billion after its $65 billion Series H, confidentially filed for its IPO this month. → thenextweb.com

Synthszr Take: 29 million exchanges via 25,000 fake accounts is no accident; it's industrial-scale replication. The economics behind it are fascinating: Anthropic burns billions in compute and training, and a distiller approximates the result for a fraction of the cost. This is the Jevons paradox of model IP: the more valuable the engineering skills, the more aggressively they get siphoned. We saw this back in February when DeepSeek and others were named; now a $965 billion company is heading for an IPO and suddenly needs an IP fence around software that is, by definition, copyable. This is where it gets Kafkaesque: the same US government that is supposed to slow down China labs is simultaneously blocking Anthropic's own models for foreign users via the Lutnick Order. Anyone who builds their moat purely on model advantage has a moat made of water. The lesson for any engineering org betting on a single frontier model: the sustainable advantage lies in deployment and knowledge diffusion within the team, not in the weight of the model that will be running, distilled, in a competitor's cloud tomorrow.

Computer Use Moves into the Base Model: Gemini 3.5 Flash Now Controls the Screen Itself

Yesterday, Google built the Computer Use capability directly into Gemini 3.5 Flash. What was previously a standalone model (Gemini 2.5 Computer Use) is now natively integrated into the main model and, according to Google, delivers the best performance to date for agentic computer tasks. This allows developers to build agents that can see, reason about, and operate browsers, mobile, and desktop environments, for tasks like continuous software testing or knowledge work across multiple applications. To combat prompt injection, Google relies on targeted adversarial training and two optional enterprise protection systems: explicit user confirmation for sensitive actions and automatic task termination upon detected injection. Access is available through the Gemini API and the Gemini Enterprise Agent Platform, with demos hosted by Browserbase. Google recommends a defense-in-depth approach, including sandboxing, human-in-the-loop, and strict access rights. → blog.google

Synthszr Take: The real leverage is in the word 'Flash.' Computer Use is no longer an expensive specialized model but part of the cheap, fast standard model, and that changes the calculus for anyone planning large-scale automation. It's reminiscent of the leap triggered by Claude Code in late 2025: the AI is leaving the chat window and reaching into where the work gets done. For a German company with 200 engineering roles, this means that nightly refactorings, software testing, and knowledge work across SAP, Salesforce, and Office suddenly run at token prices that fit on a CFO's dashboard. The sore spot remains prompt injection in live environments, and Google's two protection systems are a start, but no one should let an agent loose on production systems without a sandbox and access control. Those who don't start preparing their own operating streams for such agents now will pay more for the learning curve later. The question is no longer if, but which three use cases go into the pipeline tomorrow morning.

01.AI Wants to Become China's Palantir

Kai-Fu Lee has almost silently restructured his AI startup 零一万物 (01.AI) over the past year, meeting with over a hundred CEOs in that time, half from China and half from abroad. Instead of focusing on expensive foundation models, which have become a pure matter of capital in the token price war following DeepSeek, the company is now explicitly modeling itself not on OpenAI or Anthropic, but on Palantir: data integration, decision support, and execution systems for governments and key industries. Lee is avoiding the US market and instead traveling to Central Asia, Southeast Asia, the Middle East, Europe, and Africa. In Kazakhstan, he sits on the country's AI development council, and with the agricultural giant Charoen Pokphand, he is building the joint venture 万蜂智能, which sends AI to chicken farms to improve mortality and margins. The order numbers reflect this shift: around 500 million yuan in 2025, with 1.5 billion yuan in contracts already signed for 2026, plus a new financing round and a planned IPO in 2027. Lee's core thesis is that real contracts aren't for 200,000-yuan service agents, but for a complete AI transformation that only the '一号位' (yī hào wèi), the number one person in a company, can decide on. → Hello China Tech

Synthszr Take: Lee is making the only move left for a 65-year-old with no appetite for a capital war, and he's doing it smartly. While the other '大模型六小虎' (dà móxíng liù xiǎo hǔ - The Six Little Tigers of Large Models) burn billions to build a Chinese OpenAI, he goes where the margins are: onto the client's balance sheet. His best line is the diagnosis that treating an AI project as an IT initiative is like putting a locomotive engine in a horse-drawn carriage. This aligns with what we described in March in 'CODE CRASH': when coding and models become an interchangeable commodity, the value shifts from the model to the transformation of the organization around it. Skepticism remains about sovereignty AI as a business model, as presidential meetings in Astana and joint ventures on chicken farms scale less well than a SaaS product, and the Palantir comparison carries a political risk that Lee sugarcoats. But the direction is right: it's better to price your fee below McKinsey and deliver results instead of PowerPoint than to keep flogging the dead horse of the foundation model race. Anyone who collects 1.5 billion yuan in contracts before announcing an IPO has understood that viability comes before vision.

L'Oréal is Writing the Training Data Itself

An entire industry has specialized in writing brands into the answer machines from the outside: Generative Engine Optimization, prepared Reddit threads, reverse engineering what the model reads. L'Oréal is skipping all that and giving OpenAI its product database directly. Maybelline's virtual try-on no longer works via the detour of optimization but sits inside the system. If you have a direct line to OpenAI, you no longer have to tug at the training data from the cheap seats. L'Oréal answers the question “How do I influence what the model knows about me?” by becoming the source itself. Next to this, GEO looks like SEO for everyone who doesn't have a number in OpenAI's address book. → AI Secret

Synthszr Take: In Code Crash, it says that web mentions correlate about three times more strongly with AI visibility than backlinks (0.664 vs. 0.218), and that the model doesn't believe the 'About Us' page, but what others say. L'Oréal has turned the tables and is becoming what the model reads. This is the real moat in this phase: not the clever tactic from the outside, but the direct seat at the source. Whoever supplies the database determines the 'share of model' in their category before the GEO consultants have even opened their checklists. This is exactly why the competition is shifting from “Who optimizes better?” to “Who gets invited to co-write?”. For mid-sized brands without a direct line to OpenAI, this means finding the unoccupied fields that the big players haven't cleared yet, and occupying them while the model still doesn't know them. The whitespace in the model is closing faster than most marketing budgets can react.

At Meta, the Agent Writes the Status Report, Not the Team

Nikhyl Singhal spoke with Jagjit Chawla, a VP of Product at Meta responsible for Feed, Reels, and now Search, on the Skip Podcast. The key takeaway: once ideas become cheap, judging ideas becomes the real job, and the PM becomes the biggest bottleneck. For his teams, the classic PRD has shrunk to one paragraph plus a prototype, supplemented by an eval set in the ML domain. Status reports are no longer sent up the hierarchy via decks: Chawla's agent reads every code diff, every email, and every document at night and presents him with a bulleted list at 7 a.m., project by project in red, yellow, and green. In parts of his organization, the agent even writes the review document, highlights the three contentious questions, and names the five people needed in the room. A data query that used to require a 24-hour handoff to data science is now answered by an analytics agent in five minutes. All of this was built not by central mandate or procurement, but by the teams themselves, sometimes in a single evening. → Lenny's Newsletter

Synthszr Take: What Chawla calls the 'compression algorithm' is exactly the shift I describe in Code Crash: previously, code was expensive and intent was cheap; now it's the other way around. When output costs nothing, input becomes the scarce resource, and judgment beats production. What's interesting isn't so much that Meta is using AI (everyone is doing that right now), but that the pipeline is being radically condensed: the review starts with the decision instead of thirty minutes of context building. For decades, the org chart was the information conduit, and now a nightly agent replaces it for 50 projects simultaneously. The way it was built is remarkable. No one launched an 18-month program; the teams built their systems in an evening with tools that were already on the table. Anyone who read about the 8,000 laid-off Meta employees in May and thought it was purely about cost-cutting is seeing the other half of the equation here: the work itself is being reinvented, and that can start tomorrow morning over coffee, not after the next strategy offsite.

Serverless Postgres for Apps and Agents

Databricks is launching Lakebase, a fully managed Postgres database that docks directly into the Lakehouse and is tailored for AI agents and apps. The trick: the database branches like code. Developers can instantly create branches for testing or development, work against production data, and roll back without touching the live environment. Compute and storage are decoupled and scale independently, with scale-to-zero when idle and pay-per-use. Agent Memory is showcased as a prominent use case: chat sessions and messages are persistently stored in Lakebase, so that an agent can access previous conversation histories across deployments. Through pgvector, PostGIS, Unity Catalog governance, and Change Data Feed, the operational storage connects to the analytics world without anyone having to build their own ETL pipelines anymore. → Latent.Space

Synthszr Take: For decades, Postgres has been the unspectacular workhorse in the engine room, and now Databricks is turning it into the memory for agents. That's the real leverage here. Agents without persistent state are forgetful interns who start from scratch after every deploy. Branching like in Git is the move that will save developers work tomorrow morning because they can test against real production data without breaking anything. The lock-in question honestly remains open: anyone who couples their operational storage to Databricks' Lakehouse will pay a migration price later, and on my vendor grid, this moves from model lock-in (low) toward platform lock-in (medium). Nevertheless, the direction is right because open Postgres with pgvector and the existing tool ecosystem (pgAdmin, DBeaver, psql) at least doesn't make the transition a dead end. Reasonable Sovereignty here means: use it, but think about the export path from day one.

ByteDance Skips Four Versions and Delivers 4K Video from 50 References

On Tuesday, at the Volcano Engine FORCE conference in Beijing, ByteDance unveiled Seedance 2.5, a model that generates 30-second clips in native 4K from a single prompt. The company jumped directly to 2.5 from its predecessor, skipping four intermediate versions to signal the leap. The core improvement is in reference capacity: up to 50 multimodal inputs (images, audio, 3D white models, style references), compared to 12 in the previous version. It also features 10-bit color depth, 20 percent better prompt adherence, native audio synchronization in the same latent space, and a 3D white-box preview for quick lofi animations before the expensive render. CEO Liang Rubo called climbing the AI summit the group's top priority. The enterprise beta is already running, with a public launch planned for early July; there is no US release date. The context is sensitive: three months ago, ByteDance had to retrofit watermarks and IP guardrails after cease-and-desist letters from Disney, Warner Bros, Paramount, and Netflix, following a viral Tom Cruise vs. Brad Pitt deepfake that triggered a formal MPA complaint. → The Next Web

Synthszr Take: The number that matters is 50 versus 3. Google's Veo 3.1 accepts three reference images; Seedance 2.5 takes fifty multimodal inputs, and it's precisely this gap that determines whether a model remains a toy for memes or moves into a professional production pipeline. While OpenAI shut down Sora in March (one million users at its peak, about a million dollars in daily operating costs, and just over two million dollars in total revenue), the Chinese providers are developing their tools into real production tools more quickly. We wrote in March that Chinese talent dominates the AI scene; here you can see what that means operationally, namely distribution via CapCut's 400 million monthly users and vertical integration from generation to editing to sharing. What ByteDance lacks is an agreement with Hollywood, and every new feature that makes the model more powerful increases the pressure in this unresolved dispute. The most honest assessment of the situation: the technology is solved, the rights issue is not, and no amount of 4K quality can compensate for a missing license. Anyone looking to act on this tomorrow should test the enterprise beta for internal pipelines and keep global availability out of the high-risk zone until the IP issue is settled.